Research on Power Load Forecasting based on Machine Learning

Accurate load forecasting is the basis to realize safe and economic running of power systems and scientific management of power grid. It is also an important guarantee to improve utilization and economic dispatching efficiency of power generation assembly. Machine learning method centered at the mathematical model which has relatively high fitting ability is conducive to effective analysis and prediction of power loads. In this study, important features of specific practical hourly load data in a region provided PJM Company, such as absolute energy, approximate entropy and autoregressive coefficient, were extracted firstly by using tsfresh base in python. Practical load data of power consumers was preprocessed. Secondly, BP neural network model and XGBoost algorithm model were constructed. Based on daily practical power load data, R2 of BP neural network model and XGBoost algorithm model on the test set were tested 0.752 and 0.802, respectively. Results show that the XGBoost algorithm can fit dynamic variation trend of short-term power loads more accurately, thus enabling to increase prediction accuracy of the short-term power loads. Hence, the XGBoost algorithm possesses promising application prospects.